Cascaded Channel Estimation for Large Intelligent Metasurface Assisted Massive MIMO

• Massive MIMO supports high data traffic and many wireless connections, but large-scale arrays create high hardware costs and power consumption.Source: §I · p. 1
• LIMs are proposed as energy-efficient surfaces that manipulate electromagnetic waves through controllable reflection for wireless communications.Source: §I · p. 1
• CSI for both BS-LIM and LIM-receiver links is essential for reflect beamforming, energy-efficient design, and simultaneous passive beamforming and information transfer.Source: §I · p. 1
• Because the LIM only passively reflects waves without signal processing, cascaded estimation is formulated as sparse matrix factorization followed by matrix completion.Source: §I · p. 1
• The paper claims the first cascaded-channel estimation attempt for LIM-assisted MIMO with all LIM elements passive, unlike prior work using a few active elements.Source: §I · p. 1

• The LIM has N low-cost passive elements, the BS has M transmit antennas, and each user has L receive antennas.Source: §II · p. 2
• The direct BS-user link is neglected or estimated separately and cancelled, while channels remain unchanged during each coherence block of length T.Source: §II · p. 2
• Each LIM element uses an on/off state and a phase shift, enabling programmable passive beamforming based on environmental changes.Source: §II · p. 2
• The block-fading model allows channel variation between transmission blocks, including cases involving node motion or millimeter-wave weather effects.Source: §II · p. 2
• The BS-LIM channel H and LIM-user channel G are assumed rank-deficient, consistent with far-field and limited-scattering massive-MIMO models.Source: §II · p. 2

• The objective is to estimate cascaded channel matrices H and G from Y when pilot symbols X and S are known at the receiver.Source: §III · p. 2
• Knowledge of H and G supports reflect beamforming, energy-efficient design, and simultaneous passive beamforming and information transfer.Source: §III · p. 2
• Constant LIM phase-shift vectors within a block yield effective channels G′ and H′ related by an invertible diagonal matrix Φ.Source: §III · p. 2
• Training must therefore estimate G and H up to the diagonal ambiguity induced by Φ through appropriate designs of LIM and BS pilots.Source: §III · p. 2

• Constant LIM states do not provide sufficient structure for cascaded estimation because the resulting problem is affine matrix factorization.Source: §IV · p. 2
• The received signal is recast as a bilinear model with Z = S ⊙(HX), and the method estimates G and Z before recovering H.Source: §IV · p. 2
• Random Bernoulli on/off LIM pilots expose partial samples of Z, while the transmitted pilot X must be full-rank.Source: §IV · p. 3
• The sparse-factorization stage recovers G and Z using techniques such as BiG-AMP, while the matrix-completion stage fills missing Z entries using H's rank deficiency.Source: §IV · p. 3
• The JBF-MC algorithm retains the diagonal ambiguity because it does not need to be removed for the beamforming-related objective.Source: §IV · p. 3

Sparse Matrix Factorization

• BiG-AMP approximately computes MMSE estimates of G and Z from Y using marginal posterior means.The algorithm uses sum-product loopy belief propagation with approximate marginal posteriors.Source: A. Sparse Matrix Factorization · p. 3
• The method uses independent Gaussian priors for G and independent Bernoulli-Gaussian priors for Z.The prior parameters include the average variances ν_g and ν_z.Source: A. Sparse Matrix Factorization · p. 3
• The JBF-MC algorithm takes Y, S, X, and prior distributions p(G) and p(Z) as inputs before sparse factorization and matrix completion.Source: A. Sparse Matrix Factorization · p. 3
• Sparse matrix factorization has a diagonal ambiguity: (ĜΦ, Φ^-1Ẑ) is also valid when Φ is an invertible diagonal matrix.This ambiguity motivates using matrix completion to recover the remaining channel information.Source: A. Sparse Matrix Factorization · p. 3

Matrix Completion

• RGrad solves a matrix completion problem to recover missing entries of Ẑ using the rank-deficient property of H.The matrix completion procedure is summarized in Lines 25 to 32 of Algorithm 1.Source: B. Matrix Completion · p. 4
• The algorithm projects updates onto the left singular-vector subspace associated with the first r eigenvalues of the current estimate.Source: B. Matrix Completion · p. 4
• The hard-thresholding operator H_r retains the best rank-r approximation obtained from the singular value decomposition.Source: B. Matrix Completion · p. 4
• After RGrad outputs Â, the estimate of H is computed using ÂX†, assuming T ≥ M and rank(X) = M.These conditions ensure that X† exists.Source: B. Matrix Completion · p. 4

Computational Complexity

• BiG-AMP requires at most I_maxJ_maxO(LNT) flops because its dominant operations are basic matrix multiplications.Source: C. Computational Complexity · p. 4
• RGrad costs at most K_maxO(rNT) flops, using r < min{N,T}.Its dominant per-iteration operations require O(rNT) and O(r^3) flops.Source: C. Computational Complexity · p. 4
• Computing the final H estimate requires O(M^2T) + O(M^3) + O(MNT) flops.Source: C. Computational Complexity · p. 4

Simulation Setup

• The simulations use half-wavelength ULA configurations at the BS, LIM, and receiver, with channel matrices generated by superposing propagation paths.Source: §V · p. 4
• The BS-LIM channel uses few enough paths to have a low-rank structure, facilitating its estimation in the matrix completion stage.Source: §V · p. 4
• NMSE is averaged over 200 independent trials, with SNR defined as 10 log10(1/σ^2) dB.Diagonal ambiguities in the outputs are removed using the true G and H values when calculating NMSE.Source: §V · p. 4

Simulation Results

• Under λ = 0.2, JBF-MC has a significant performance gain over baseline methods, especially for estimating G, while H has lower NMSE than G.Figure 2 varies SNR and pilot number with N = 70 and M = L = 64.Source: §V · p. 4, Figure 2 · p. 5
• The sampling rate has a tradeoff: too small causes matrix completion to fail, while too large causes sparse matrix factorization to fail.Larger λ supplies more matrix-completion samples but increases the random variables estimated during BiG-AMP.Source: §V · p. 5, Figure 3 · p. 5

• The framework combines bilinear sparse matrix factorization with matrix completion for cascaded channel estimation.Source: §VI · p. 5
• The proposed two-stage algorithm includes generalized bilinear message passing for matrix factorization and a Riemannian manifold gradient-based algorithm for matrix completion.Source: §VI · p. 5
• The proposed approach achieves accurate cascaded channel estimation for LIM-assisted massive MIMO systems.Source: §VI · p. 5